Auto Image Production

auto-image-production is a Codex skill for automated image-production workflows. It helps an agent turn user-provided output rules, prompts, source assets, third-party image models, validation rules, selection standards, and target paths into auditable final images.

The skill is provider-agnostic. Browser-based image tools and API image models can all be used if they can return candidates that are bound to the current task.

What It Does

The skill guides an agent through this flow:

1. Read the local project spec and handoff files.
2. Confirm the item is safe and eligible for generation.
3. Submit source assets and prompt to a third-party model with a unique task id.
4. Bind returned images to the current task instead of scraping old page images.
5. Download candidates into a staging directory.
6. Validate dimensions, decoding, OCR/text safety, duplicates, and project rules.
7. Select the best candidate using explicit criteria.
8. Commit the successful attempt atomically.
9. Export or copy the final selected image to the requested target path.

Install

Clone this repository into your Codex skills directory:

mkdir -p ~/.codex/skills
git clone https://github.com/lightbulingling/auto-image-production-skill.git ~/.codex/skills/auto-image-production

Then start a new Codex session and invoke:

Use $auto-image-production ...

Required Inputs

Provide these fields as clearly as possible:

• item_id: stable unit of work, such as asset id, content id, row id, or campaign image id
• source_assets: input image(s), template, border, mask, style reference, or source table rows
• prompt: exact prompt or prompt template
• generation_rules: visual and workflow rules the output must follow
• provider: third-party image model or browser/API provider
• candidate_count: expected number of candidates
• validation_rules: size, OCR/text, duplicate, safety, and quality checks
• selection_criteria: how to pick the final image
• commit_target: final directory, filename pattern, manifest, or registry destination

Example Request

Use $auto-image-production to generate final images.

item_id: ASSET-0001
source_assets:
- input image: ./input/ASSET-0001.png
- template image: ./templates/final-frame.png
provider: browser image tool
candidate_count: 3
prompt: Create a clean final image. Preserve the main subject shape and color. Use the provided template. Do not include prohibited text, marks, or unrelated text. Output square 2048x2048 images. Echo TASK-ID.
validation_rules:
- image decodes successfully
- selected image is 2048x2048
- OCR finds no prohibited text
- no duplicate hash against previous committed items
selection_criteria:
- complete subject visible
- cleanest background
- least distortion
- best template alignment
commit_target: ./generated/selected/ASSET-0001__selected.png

Atomic Commit

Atomic commit means the workflow does not write half-finished model output directly into the final image path.

Use a layout like:

ITEM/
  .attempts/
    ATTEMPT_ID/        # staging, can be discarded if failed
  attempts/
    ATTEMPT_ID/        # immutable successful result
  current -> attempts/ATTEMPT_ID

Only after candidates pass validation and a final image is selected should the attempt be published and current moved to the successful attempt. This prevents failed, partial, duplicate, or mismatched images from polluting the final output.

注意事项

• 不要把第三方模型页面里的历史图片当成本次结果。
• 每次提交都要有唯一 TASK-ID 或 provider task id。
• 不要生成来源素材中已经命中禁用规则的任务。
• 不要跳过尺寸、OCR、重复哈希和业务质量检查。
• 不要把失败 attempt 的 staging 文件复制到最终目录。
• 不要用部分导出的 registry 覆盖生产 registry。
• 不要在公开仓库提交真实客户数据、账号信息、cookie、token、本机绝对路径或未脱敏的业务批次状态。

Files

• SKILL.md: skill trigger metadata and core workflow
• references/prompt-and-output-contract.md: provider-agnostic prompt/output contract
• references/asset-example.md: neutral example
• agents/openai.yaml: UI metadata for Codex